Executive Summary
Israel carried out strikes in Tehran early Saturday, and a U.S. official confirmed American participation, a sharp escalation that landed amid fragile U.S.-Iran nuclear diplomacy and a growing American military buildup across the region. Hours earlier, Pakistan’s defense minister declared “open war” on Afghanistan after airstrikes hit Kabul and Kandahar, with both sides issuing sweeping casualty claims that could not be independently verified. In Washington, the Pentagon’s showdown with Anthropic over A.I. safeguards intensified ahead of a Friday deadline, even as OpenAI disclosed a colossal funding round and a deepened partnership with Amazon that would lock in years of cloud spending. Across Asia and Europe, governments moved to fortify strategic supply chains and deterrence postures, from Japan’s new financing push for next-generation chips to Nordic leaders signaling openness to hosting NATO nuclear weapons.
AI & Technology
Pentagon Threatens to Blacklist Anthropic in Safeguards Standoff
The Pentagon has threatened to cut ties with Anthropic and potentially label the company a “supply chain risk” unless it loosens restrictions on how its Claude models can be used, setting a Friday deadline that has drawn lawmakers into what is becoming a test case for the limits of private-sector control over military A.I. According to multiple reports, Defense Secretary Pete Hegseth demanded Anthropic agree to permit “all lawful uses” of its technology, language the department argues is necessary for operational flexibility and the protection of U.S. forces.
Dario Amodei, Anthropic’s chief executive, has refused. He has said the company can support specific defense missions but cannot approve uses that could enable mass domestic surveillance or fully autonomous weapons without meaningful human control. Pentagon officials, in turn, have argued that determining legality is the government’s role, not a contractor’s, and have accused Mr. Amodei of exaggerating the department’s intentions; one official described his stance as a “God-complex,” according to accounts of the dispute.
The confrontation has quickly become political. Senators from both parties have urged the Pentagon to withdraw its deadline and sought legislative oversight, warning that the military is effectively trying to set national policy by contract. The Pentagon’s willingness to invoke tools like the Defense Production Act, and its talk of “blacklisting,” signals how aggressively it intends to secure access to frontier models, even at the risk of chilling cooperation across the sector.
It is unclear whether Anthropic will hold its line if the department follows through, or whether the Pentagon would accept narrower language that preserves the company’s red lines while giving commanders practical latitude. Either way, the clash is already forcing a more public reckoning over a question the defense establishment has largely tried to settle quietly: who decides what “lawful” means when a model’s capabilities can be repurposed faster than statutes can be written.
OpenAI’s $110 Billion Raise Deepens Reliance on Amazon’s Cloud
OpenAI has closed a funding round of up to $110 billion that values the company at roughly $730 billion, while announcing a far-reaching partnership with Amazon that would bind one of the world’s most prominent A.I. developers to one of the world’s largest cloud providers. The deal includes a $50 billion investment by Amazon and a commitment by OpenAI to spend $100 billion over eight years on Amazon Web Services computing and chips, on top of an existing $38 billion contract, according to reports describing the arrangement.
Nvidia and SoftBank are said to have committed $30 billion each. Microsoft, OpenAI’s largest shareholder and long-time partner, did not participate, though OpenAI said the financing did not alter Microsoft’s license and access to OpenAI’s intellectual property. Still, the absence was striking in an industry where capital and compute have become inseparable, and where strategic investors increasingly shape the direction of model development.
The companies said they would jointly develop custom A.I. models for Amazon’s consumer products and offer a “Stateful Runtime Environment” powered by OpenAI models to AWS customers through Amazon Bedrock. AWS would also become the exclusive third-party cloud distribution provider for OpenAI’s enterprise platform, Frontier, giving Amazon a privileged position in delivering OpenAI’s tools to corporate clients beyond OpenAI’s own channels.
Investors are treating the round as a wager that OpenAI can keep pushing the frontier fast enough to justify a valuation that would have seemed implausible even two years ago. Critics, including some market analysts, have warned that the industry’s largest players are stacking long-term compute obligations on top of enormous cash raises, creating a cycle in which the next model’s cost requires the next round of financing. OpenAI has floated an initial public offering as soon as the end of the year, but the company’s ability to reconcile its ambitions with its expanding web of contractual commitments will be one of the questions that determines whether that timeline holds.
Meta Rents Google’s A.I. Chips, Undercutting Nvidia’s Grip
Meta has struck a multibillion-dollar deal to rent Google’s Tensor Processing Units, according to The Information and Reuters, a move that would deepen an unlikely alliance between two rivals while reshuffling the power dynamics of the A.I. chip market. The arrangement focuses on Google’s proprietary TPUs, which were originally developed for internal workloads and then gradually commercialized through Google Cloud, and could expand to direct purchases by Meta as early as next year.
The deal arrives as Meta spends aggressively on compute, including a reported agreement with AMD that could reach $60 billion and expanded purchases from Nvidia, while building a $10 billion data center in Lebanon, Ind. For Google, renting TPUs at scale is a way to turn one of its key competitive advantages—specialized inference hardware—into a revenue engine and a lever of influence over the broader A.I. ecosystem.
The news also revived jitters about Nvidia’s dominance. Reports have suggested that earlier disclosures of Meta’s shifting mix of chip suppliers contributed to a roughly $250 billion drop in Nvidia’s market value in December, an illustration of how sensitively investors now react to any hint that Big Tech might diversify away from one supplier.
Elsewhere in the infrastructure race, MARA Holdings, the Bitcoin miner, said it would partner with Starwood Property Trust to convert some mining sites into A.I.-focused data centers, using existing access to low-cost power and grid interconnections. The pivot reflects a broader pattern: as A.I. demand strains electrical capacity, firms that already control energy-heavy real estate are trying to reposition themselves as suppliers of “power-to-compute,” even if it is unclear how quickly those sites can be retrofitted to meet the reliability and cooling standards frontier-model operators require.
Japan Backs Rapidus With $1.7 Billion for 2-Nanometer Chips
Japan’s Rapidus, founded in 2022 with the ambition of restoring the country’s standing in advanced semiconductor manufacturing, has secured roughly $1.7 billion in new financing to pursue mass production of 2-nanometer chips by 2027. The funding includes about ¥100 billion from the Information-technology Promotion Agency and roughly ¥167.6 billion from 32 companies, including Canon, Fujitsu and Sony, as well as backing tied to institutions such as the Development Bank of Japan.
Rapidus is collaborating with IBM and the Belgian research center Imec, leaning on outside expertise in a field where the technical barriers are punishing and the incumbents—Taiwan Semiconductor Manufacturing Company and Samsung—have spent decades and hundreds of billions of dollars building manufacturing ecosystems. The compressed timeline is part of the point: Japanese officials see semiconductors not just as an economic sector but as a national security dependency exposed by pandemic-era shortages and intensifying U.S.-China technology rivalry.
The financing comes amid continued strains in the broader chip supply chain, particularly for memory components increasingly pulled into A.I. workloads. Consumer electronics companies have warned of delays and potential price increases, with the constraints showing up in everything from gaming consoles to the parts suppliers that feed them. The shortages have become a reminder that even as the industry races toward 2-nanometer production, large segments of the market remain vulnerable to bottlenecks in older but essential components.
Whether Rapidus can hit its 2027 target is uncertain. Japan has attempted semiconductor revivals before, only to struggle against scale economics and the gravitational pull of established foundry networks in Taiwan and South Korea. But the geopolitical premium on supply resilience, and the sheer volume of capital now pouring into A.I.-linked demand, is giving long-shot industrial bets more room to run than in earlier cycles.
Geopolitics & Security
Israel Hits Tehran as U.S. Confirms Participation
Israel struck Iran early Saturday, with explosions reported in central Tehran and Iran closing its airspace, an escalation that Israeli officials framed as preemptive and American officials acknowledged as a joint operation. Israel Katz, the defense minister, said the strikes were meant “to remove threats to Israel,” and the Israeli military issued a “proactive alert” warning citizens to stay near shelters in anticipation of Iranian retaliation.
A U.S. official told Al Jazeera that the United States participated in the attack, a confirmation that brings Washington directly into a confrontation it has often sought to keep at one remove. The explosions were reported near offices associated with Iran’s supreme leader, Ayatollah Ali Khamenei, though it was unclear whether he was present. Iranian state television acknowledged an explosion but provided few details, a familiar pattern in moments when Tehran is calibrating its public messaging as carefully as its military response.
The strike comes as indirect U.S.-Iran nuclear talks—already strained—move toward technical discussions in Vienna next week, after a round in Geneva ended without an agreement. President Trump said on Friday that he was “not happy” with the trajectory of the negotiations and demanded a “meaningful” deal to ensure Iran does not develop nuclear weapons. In separate remarks, he suggested force remained an option, saying, “sometimes you have to,” while also insisting he preferred to avoid conflict.
American military posture had been visibly hardening even before Saturday’s attack. The United States has repositioned major assets in the region, including the aircraft carrier USS Gerald Ford off Israel’s coast, and authorized the departure of non-emergency personnel and family members from the U.S. Embassy in Jerusalem. The U.S. Embassy in Qatar issued shelter-in-place guidance for its personnel, underscoring concerns that Iran might retaliate beyond Israel’s borders, potentially against American bases within range of Iran’s missile arsenal.
Iran has options that fall short of a direct state-to-state exchange—proxy attacks, maritime disruption, cyber operations—yet the U.S. confirmation of participation narrows the space for Tehran to respond without escalating against U.S. forces. It is unclear whether the strikes were designed to degrade specific nuclear or missile capabilities, to shape negotiations, or to force a broader strategic reordering. What is clearer is that any Iranian retaliation will now land on a region already braced for cascading consequences.
Munition Shortages Shadow Trump’s Iran Options
As the administration weighs next steps on Iran, senior military briefings have highlighted a constraint that rarely dominates public debate but can shape war plans as decisively as politics: the supply of defensive interceptors. Adm. Brad Cooper, the commander of U.S. forces in the Middle East, and Gen. Dan Caine, the chairman of the Joint Chiefs of Staff, briefed Mr. Trump on Thursday on possible military actions against Iran, according to reports, against the backdrop of nuclear diplomacy that some officials described as a final push before decisions harden.
The concern is that a large Iranian missile and drone barrage could consume vast numbers of interceptors—particularly THAAD munitions—at a rate that would strain U.S. readiness elsewhere. Analysts cited in the reports said the U.S. and Israel used a substantial number of interceptors during a recent 12-day conflict in which Iran launched hundreds of missiles. One estimate put U.S. THAAD usage at up to 150 interceptors in defense of Israel, out of fewer than 650 acquired since 2010.
That arithmetic has strategic implications. If U.S. planners expect Iran to retaliate at scale, they must weigh how many interceptors can be spared for the Middle East without weakening stocks needed for Ukraine, or for contingencies involving China or Russia. The effect could be to press Israel to shoulder more of its own air defense burden, while the United States prioritizes protecting its forces and critical regional infrastructure.
The administration’s rhetoric has left room for multiple theories of the case. Mr. Trump has alternated between describing diplomacy as viable and emphasizing the capacity to strike Iran’s nuclear infrastructure, referencing deployments including B-2 bombers. Critics have argued that the messaging has been inconsistent, particularly after earlier operations against Iranian facilities that the administration described as decisive while downplaying dissenting assessments.
For now, the diplomatic track is still moving: consultations in Washington and Tehran are expected to precede technical discussions in Vienna. Yet the operational reality described in the briefings suggests that even limited military actions could impose longer-term costs, not only in escalation risk but in the mundane, determinative question of how quickly factories can replenish what missile defenses expend in days.
Pakistan Bombs Kabul and Kandahar, Declaring “Open War” on Taliban Rule
Pakistan launched airstrikes on Kabul, Kandahar and other Afghan locations on Friday, prompting Islamabad’s defense minister, Khawaja Mohammad Asif, to declare “open war” on Afghanistan’s Taliban-led government after days of cross-border clashes. Pakistan’s military said it struck 22 sites and claimed that more than 200 Taliban fighters were killed. The Taliban, in a competing account, said Pakistan’s attacks killed 13 Taliban fighters and 13 civilians, and asserted that Afghan forces captured 19 Pakistani military posts and two bases, killing 55 Pakistani soldiers. None of the casualty figures could be independently verified.
The airstrikes marked one of the most serious confrontations between the neighbors since the Taliban returned to power in 2021, and among the most expansive bombardments of Kabul in years. Pakistani officials have long accused the Taliban government of sheltering militants from Tehrik-i-Taliban Pakistan, the Pakistani Taliban, and other insurgent groups that Islamabad says have fueled rising attacks since 2022. Kabul has rejected those claims, arguing that Pakistan’s security problems are homegrown, and counter-accusing Pakistan of harboring fighters from the Islamic State.
Mr. Asif portrayed the conflict in sweeping geopolitical terms, saying Pakistan’s “cup of patience has overflowed” and accusing the Taliban of turning Afghanistan into a “colony of New Delhi,” an allegation India has historically denied. His remarks, which also criticized the Taliban’s human rights record, suggested that Islamabad may be trying to justify a sustained campaign, not a single punitive raid, even as it insists there will be “no dialogue” until cross-border militancy stops.
Afghan officials said they had shot down a Pakistani F-16, a claim Pakistan has not confirmed and independent analysts have disputed. If the Taliban’s assertion were substantiated, it would raise uncomfortable questions for Pakistan’s air operations and for the Taliban’s actual air-defense capabilities, which are widely seen as limited.
International actors including the United Nations and the European Union urged restraint, while regional powers—among them Iran, Russia, Jordan and the United Arab Emirates—called for de-escalation. But the fighting has already widened beyond conventional engagements. A drone attack on a mosque in Bannu, Pakistan, injured at least five people, according to reports, and Afghan outlets described Taliban drone attacks on Pakistani camps. With both sides claiming territorial gains along a 2,600-kilometer border, the immediate question is whether this becomes a grinding frontier conflict or a crisis that outside mediators can still contain.
Indonesia Plans Gaza Deployment as Nordic Leaders Reopen Nuclear Debate
Indonesia is preparing to send about 1,000 soldiers to Gaza within weeks as part of a U.S.-led International Stabilization Force, according to reports, a deployment that would be the first contingent of an 8,000-strong mission expected to grow through June. The plan has stirred debate inside Indonesia, the world’s largest Muslim-majority nation, where critics fear the force could be used to legitimize Israeli control rather than protect Palestinians.
Shofwan Al Banna Choiruzzad, an associate professor at the University of Indonesia, warned that without United Nations oversight, Indonesia’s troops could become “pawns” of Washington and Israel. Supporters of the deployment have pointed to Indonesia’s long record in UN peacekeeping and argued that a stabilizing presence could relieve humanitarian pressure. But details of the mission’s mandate and command structure remain murky, and it is unclear how rules of engagement would function in an environment where the line between stabilization and enforcement is often contested.
At the same time, security debates in Europe continued to harden. Sweden and Denmark have signaled openness to hosting NATO nuclear weapons, a striking shift for countries that only recently joined the alliance and long treated nuclear basing as politically toxic. Sweden’s defense minister, Pal Jonson, said the country would consider “any option that could secure Sweden’s survival” in a war, while Denmark’s defense minister, Troels Lund Poulsen, expressed similar openness amid broader European discussions about deterrence and defense spending.
These moves unfolded alongside a growing U.S. military buildup around Israel and the Gulf, including carrier movements and guidance for U.S. diplomatic families to depart. The developments are not directly linked, but they reflect a shared logic of the moment: governments are repositioning forces and rewriting old taboos in response to threats that feel less hypothetical than they did a few years ago. What remains unclear is whether these steps will deter escalation—or, by normalizing higher-risk postures, make crises harder to defuse when they arrive.
Economy & Markets
The A.I. Buildout Spreads From Chips to Power, Testing Regulators
The week’s biggest A.I. deals underscored how the industry’s economic center of gravity has shifted from software to infrastructure, with cloud commitments, specialized chips and power-hungry data centers now driving strategy as much as model performance. OpenAI’s agreement to spend $100 billion on AWS over eight years and Meta’s reported move to rent Google’s TPUs point to a market in which access to compute can dictate product road maps and competitive positioning.
For Nvidia and other chipmakers, the pressure is twofold. On one hand, demand remains extraordinary; on the other, Big Tech is actively building alternatives, whether through Google’s commercialization of TPUs or Meta’s willingness to diversify suppliers across Nvidia and AMD. Investor volatility has followed, with reports tying prior news of Meta’s shifting chip mix to a steep single-month decline in Nvidia’s market value.
The expansion is also pulling in unlikely players. MARA’s effort to convert Bitcoin mining sites into A.I. data centers reflects the value of existing energy and grid access, even as communities and regulators scrutinize power use, water consumption and the broader footprint of large-scale compute facilities. The conversions may appeal to financiers because they turn a volatile crypto-revenue model into contracted infrastructure income, but operators still face the challenge of delivering the uptime and cooling capacity that A.I. workloads demand.
Regulators in the United States and Europe have been watching these alliances with increasing intensity, not least because they blur the line between competition and collaboration among the largest platforms. When rivals rent chips to rivals and lock in cloud distribution exclusivity, antitrust questions tend to follow—especially as governments also press the same firms to provide national security capabilities. The result is a market where commercial concentration and strategic dependency are colliding, and where the next set of rules may be written as much in procurement offices as in legislatures.
Regional Developments
Xi Purges Officials Ahead of Congress as Beijing Downshifts Growth Targets
China has removed 19 officials—including nine military figures—from its list of lawmakers ahead of the National People’s Congress, a move that appears to extend President Xi Jinping’s anti-corruption campaign deeper into the armed forces and the political class. Among those removed were senior officers including Li Qiaoming, the commander of the People’s Liberation Army Ground Force, and Shen Jinlong, a former navy commander, according to reports. No official reasons were provided, and the lack of explanation has fueled speculation that the reshuffle is as much about power consolidation as discipline.
The personnel changes come as Beijing signals a notable reorientation of economic policy. In a recent Politburo meeting chaired by Mr. Xi, the leadership emphasized domestic demand and “new growth drivers,” while calling for “effective improvement in the quality of the economy and a reasonable growth in its quantity.” The language has been read as a rebuke to the growth-at-all-costs incentives that have long shaped local governance, even as deflationary pressures, property-sector weakness and subdued consumption weigh on the economy.
Local officials, accustomed to promotion systems tethered to headline GDP, have reportedly been left uncertain about what metrics will replace the old ones. The central government has ordered party members to correct “deviations” in their understanding of performance, an admonition that suggests compliance will be enforced politically as well as administratively.
The National People’s Congress is expected to set a lower growth target for 2026—possibly between 4.5 and 5 percent—alongside new guidance on employment, debt and risk management. Whether the combination of political tightening and economic recalibration produces a more resilient model, or simply deepens caution within the bureaucracy, will shape not only China’s trajectory but the global trade and investment environment tied to it.
Beijing Courts Germany and Watches Taiwan’s Rhetoric Ahead of Xi-Trump Summit
China has been managing a complicated set of relationships as it prepares for a possible summit between Xi Jinping and President Trump, while also seeking to stabilize ties with Europe and reduce cross-strait friction. German Chancellor Friedrich Merz visited China recently, emphasizing economic ties even as Berlin and Brussels remain wary of trade imbalances and dependence on Chinese supply chains.
In Taiwan, President William Lai Ching-te has adopted more cautious language about “mainland China,” a shift analysts interpret as an attempt to lower the temperature ahead of the anticipated Xi-Trump engagement. Any substantive breakthrough on Taiwan is widely viewed as unlikely, but even modest changes in tone can affect military signaling in a region where misreading intentions can have outsize consequences.
China’s regional dilemmas extend beyond Taiwan. In Myanmar, the conclusion of junta-staged elections has done little to clarify Beijing’s long-term approach to an unstable neighbor that sits astride critical border trade routes and infrastructure corridors. China has tried to balance relationships with the military authorities and various armed groups, but the conflict’s durability has made clean solutions elusive.
Together, the diplomatic choreography suggests a Beijing trying to widen its room for maneuver at a moment when U.S. pressure is rising and economic headwinds at home are persistent. Whether that strategy yields tangible concessions—or simply buys time—will depend on choices in Washington, Taipei and European capitals as much as in Beijing.
From the Timeline
AI labs vs. national security: “reasonable agreements” or an un-adoptable rug-pull risk?
@sama framed the newly announced deployment agreement with the “Department of War” as a safety-forward template—explicitly rejecting domestic mass surveillance and autonomous weapon use while still entering classified environments. In contrast, @chamath argued the Anthropic standoff creates an adoption hazard for any enterprise or government customer: if ToS boundaries can shift with leadership or politics, reliability becomes the product, and “no holds barred” terms start to look like the safer procurement choice. The pro-principle camp leaned the other way: @ilyasut praised competitors aligning on red lines as an important precedent for tougher future scenarios, while @ClementDelangue distilled the trust debate into a simple ownership maxim (“Not your weights, not your brain”). Meanwhile, @EMostaque treated government control as the default endgame if any lab gets close to AGI—less a controversy than an expected function of state power.
“Tonight, we reached an agreement with the Department of War to deploy our models in their classified network.”
— @sama
Agentic coding reality check: speed is real, but “research orgs” still hallucinate process
@levelsio described a step-change in solo-builder throughput by running Claude Code in “dangerously-skip-permissions” mode directly on production servers—arguing the bottleneck is now personal creativity and context, not implementation time. That “go faster by loosening the leash” sentiment runs into a more cautious operational lesson from @karpathy: multi-agent “research org” setups look impressive but still produce weak experimental discipline without human steering, making them better at execution than at generating good ideas. On the measurement side, @balajis amplified forecasts that AI-authored commits could become a material share of all code soon, reinforcing the sense that the tooling shift is statistical, not anecdotal. And in the developer-experience weeds, @garrytan surfaced a quieter friction point: even top dev tools often don’t support account creation via API, a mismatch with the automation-first workflows builders are trying to adopt.
Governance backlash in the West: age gates, speech tribunals, and the “exit to friendlier jurisdictions” playbook
@dhh attacked California’s proposed OS-level age verification requirement as effectively unenforceable—another example of policymakers demanding implementation realities conform to moral urgency. On a different axis of state reach, @tobi elevated Canadian commentary arguing gender-ideology enforcement via human-rights frameworks is becoming a kind of “blasphemy law,” suggesting a broader elite legitimacy problem around compelled norms. Meanwhile, @balajis pointed to corporate migration signals (Wyoming gaining on Delaware) as the pragmatic response: founders and capital routing around jurisdictions perceived as hostile or unpredictable.
Border policy and social cohesion: competing narratives harden, and “systems are broken” becomes the shared premise
@paulg amplified reporting that appears to show Border Patrol dropping migrants outside official processing channels—fuel for critics who see a gap between stated policy and on-the-ground practice. In Europe-adjacent discourse, @elonmusk reacted to viral claims about conditions in Europe with moral outrage, framing the situation as systemic failure rather than isolated incidents. A more reformist framing showed up via @tobi, boosting an editorial that argues the refugee system is broken and that dismissing reform debates as racism is itself a way to avoid fixing incentives and outcomes. The through-line is less consensus on solutions than a convergence on “institutional trust is spent,” with each camp pointing to different culprits.
Information ecosystems: prediction markets pitch “truth,” while social feeds are blamed for rage optimization
@brian_armstrong positioned prediction markets as “info markets,” implicitly arguing that prices can outperform punditry as a public forecasting layer (even for consumer pain points like subscription price hikes). A darker media diagnosis ran in parallel: @Noahpinion amplified the idea that the internet’s engagement machine has become structurally optimized to make users angry and miserable, making “what people believe” more a product of incentive design than evidence. In a more abstract register, @pmarca argued for prioritizing better questions over definitive answers—a subtle pushback to both “the market will tell us” certainty and doomscroll fatalism.
“Prediction markets are info markets.”
— @brian_armstrong
Geopolitics and energy: alliances on display, sanctions ambiguity, and a solar-first provocation
@wolfejosh boosted India’s prime-ministerial messaging around a high-profile Israel engagement, emphasizing state-to-state “strengthened cooperation” in a tense regional context. In the Russia sanctions lane, @ylecun circulated reporting that the U.S. postponed sanctions on Lukoil’s international assets again—an implicit signal of how often enforcement timelines bend under economic and diplomatic pressure. Energy discourse stayed combative: @chamath mocked “but nuclear?!” framing by arguing solar is effectively “far-field nuclear,” pitching solar-plus-storage as the fast, cheap path that makes the traditional baseload argument feel outdated.
AI research posture: humility about “how early it is,” plus a push toward durable memory beyond context windows
@fchollet argued today’s AI moment resembles the early innings of a new science—more “1870s” than end-of-history—warning against treating current capabilities as a solved field. In a concrete research direction, @hardmaru highlighted hypernetwork-generated LoRAs (Doc-to-LoRA / Text-to-LoRA) as a path to “compile” documents/tasks into weights for fast adaptation—an attempt to move from brittle, prompt-bound working memory toward something closer to durable recall. Together, the thread reads as an implicit critique of hype cycles: capability is rising fast, but the core architecture still lacks key properties (robust generalization, stable memory) that researchers see as fundamental rather than incremental.